{"id":"cf8d78d2-d38f-42cb-8ded-a5460dce8c2d","arxiv_id":"2505.12913","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SALSA factorizes active learning over synthon choices to screen multi-vector molecular spaces up to about two trillion compounds, recovering more than 94 percent of the top-1K molecules in a one-million-molecule benchmark with a small evaluation budget.","lead":"A new algorithm, SALSA, searches huge molecule libraries built by combining fragments: it learns which fragments score well instead of scoring every molecule. It finds most of the best molecules with far fewer expensive evaluations, and it can handle spaces of trillions of compounds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SALSA's load-bearing premise—that objective scores decompose into independent synthon contributions—is only tested on ROCS-TC and docking; no interaction-dominated objective is used, so the method's general applicability remains unproven.","rationale":"The reader's weakest_assumption—independence of synthon contributions—is exactly the load-bearing point I identify. The paper's own Appendix A.4 and Conclusion concede this risk, which is why I mark agreement as 'agree' and keep the CONDITIONAL verdict. The concern is not that SALSA is invalid; the 1M-space results are an honest, enumerated benchmark showing high recall for ROCS-TC and docking, and the comparison to MolPAL is informative. But those two objectives are not enough to establish the general claim that SALSA 'extends pool-based active learning to non-enumerable spaces' for arbitrary molecular design objectives. Docking and shape similarity can be dominated by additive fragment properties (size, lipophilicity, shape volume), and the fixed core further reduces interaction complexity. A synthetic interaction-strength sweep would settle whether the method degrades gracefully or substantially when the additivity assumption is violated. The scaling experiments to 100M/10B/2T are computational demonstrations, not accuracy benchmarks: without ground truth, they cannot show near-optimality. I therefore recommend no change to the reader's CONDITIONAL verdict; the condition should be made explicit: SALSA is validated for approximately additive objectives, and its scope for strongly interacting objectives remains open pending a targeted stress test. This is a good-faith concern, not an ad hominem or a disagreement with consensus; it is a correctness risk in an unproven regime.","tokens_in":11786,"tokens_out":6329,"duration_ms":69302,"concrete_test":"Construct a synthetic 2-vector space with 1K×1K synthons and ground-truth objective f(s0,s1)=g0(s0)+g1(s1)+λ·h(s0,s1), where g0,g1 are known synthon potentials and h is a fixed random interaction matrix with entries scaled so that the interaction contributes a controllable fraction of total score variance. Compute the exact top-1K by enumeration. Run SALSA with the same budget and protocol as in Section 3 (10 rounds, K=1K, TS acquisition) for λ such that interaction variance share is 0%, 10%, 33%, and 50%. If top-1K recall degrades progressively from ~95% toward random as the interaction share grows, the independence assumption is load-bearing and the central claim must be restricted to near-additive objectives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SALSA's central claim is that factoring modeling and acquisition over synthons extends pool-based active learning to non-enumerable multi-vector spaces. The method's correctness depends on f(s0,s1) being well approximated by g0(s0)+g1(s1): the surrogate models are trained on synthon-score pairs, where each molecule's total score is assigned to both constituent synthons (Section 2, Surrogate models), and acquisition combines independently top-scored synthons. Appendix A.4 explicitly notes the resulting aleatoric variance from the unobserved complementary synthon. The paper validates this decomposition on exactly two scoring functions—ROCS-TC and OpenEye Hybrid Docking—on a single 1M-molecule CDK2 space (Section 3, Sample efficiency). Both are smooth, global functions for which fragment-level additivity may hold approximately for the chosen core; the MPO experiments (Section 3, Multi-parameter objectives) are only linear combinations of these same two scores with QED, so they add no evidence about interaction robustness. The authors acknowledge in the Conclusion that 'the implicit, naive independence assumptions ... risk breaking down against more complex objective functions.' Because the 2T scaling experiments have no ground truth, they cannot compensate for this gap. Thus the central claim holds only conditional on an additivity property that is not established for general objectives.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SALSA (Scalable Active Learning via Synthon Acquisition), an active learning method for multi-vector (R-group) combinatorial molecular spaces. SALSA trains a separate surrogate model per synthon vector on scores of complete molecules, decomposing each molecule's objective score into per-synthon contributions, and acquires new molecules by combining independently top-scored synthons. The authors report that SALSA identifies approximately 95% of the top-1K molecules in a fully enumerated 1M-molecule CDK2 space with a small scoring budget, that it scales to spaces of approximately 2 trillion molecules, and that it achieves higher scores and greater scaffold diversity than the LibINVENT generative baseline on linear multi-parameter objectives for CDK2, BACE1, and DRD2. The paper is published at the GEM workshop, ICLR 2025, and includes appendices with algorithm pseudocode, hyperparameters, ablations, and runtime tables.","tokens_in":12013,"tokens_out":4176,"duration_ms":43367,"significance":"If the reported results hold, SALSA offers a practical way to perform active learning over non-enumerable combinatorial spaces with expensive molecular objectives, which is a real bottleneck in early-stage drug discovery. The paper's strongest evidence is the honest ground-truth evaluation in a fully enumerated 1M-molecule space, with comparisons to a full-molecular active learning baseline (MolPAL) and tabular Thompson sampling, plus ablations of surrogate models and acquisition strategies. The method also provides explicit synthetic routes for proposed molecules, an advantage over many generative approaches. The principal risk is that the method's correctness rests on an additivity assumption across synthon vectors that is tested on only two objective families and acknowledged by the authors as fragile for more complex objective functions.","major_comments":[{"comment":"The central claim that SALSA generalizes to arbitrary objective functions rests on the assumption that f(s0,s1) ≈ g0(s0)+g1(s1), but the paper validates this on only two objective families—ROCS-TC and Hybrid Docking—and the MPO experiments are linear combinations of these same scores with QED. Appendix A.4 itself attributes the aleatoric variance of a synthon's score to the unobserved complementary synthon, confirming that interaction effects exist and are treated as noise. Please test SALSA on an objective with explicit R-group interactions (for example, an MPO containing a product or cross-term of the two vector contributions, or an activity model trained on both R-groups jointly) to determine whether the method remains effective when additivity is violated. If the method fails on such a task, the scope of the central claim should be narrowed, and the paper should state the conditions under which SALSA is expected to work.","section":"Section 2 (Surrogate models) and Section 3 (Sample efficiency; Multi-parameter objectives)"},{"comment":"The headline comparison to MolPAL (96.5% vs 98.5% for ROCS-TC; 94.5% vs 95.4% for docking) is reported as point means over 5 trials with no error bars, confidence intervals, or significance testing. The claimed 'minimal degradation' may be within trial-to-trial noise. Please report the per-trial variation (standard deviation, interquartile range, or confidence intervals) and, ideally, a paired test; if the difference is not statistically significant, state explicitly that the degradation is negligible rather than minimal. The same issue applies to Figure 3 (3 trials) and Figure 7.","section":"Section 3, Sample efficiency (Figure 2)"},{"comment":"The 2T-space experiments have no ground-truth ranking and no baseline (not even random acquisition), so the statement that 'SALSA consistently finds better scoring molecules with increasing space size' only demonstrates that the scores of the molecules discovered improve as the space grows, which could reflect the larger space containing intrinsically better molecules rather than SALSA's search efficiency. The reference to Lyu et al. (2023) is suggestive, but a comparison to random or frequency-based acquisition in the 10B and 2T spaces would strengthen the scaling claim; without such a control, the scaling results remain uncontrolled and should be described as proof-of-concept rather than validation.","section":"Section 3, Scaling beyond enumerable spaces (Figure 3)"},{"comment":"The ADMET comparison relies on 'internal models' to predict several properties (LogD, hERG, PXR, fraction unbound, CACO2) with no description of these models, their training data, or their accuracy. Since Figure 8 is used to claim that SALSA molecules have 'comparable chemical property profiles to known bioactives,' either report the provenance and validation of these predictors (including error bars or scatter plots against measured values) or restrict the property comparison to metrics computed directly from the molecular graph (MW, TPSA, HBA, HBD, AROM, ROTB, ALERTS).","section":"Appendix A.8 and Figure 8"}],"minor_comments":[{"comment":"Typo: 'directly combarable' should be 'directly comparable'.","section":"Section 4, Conclusion"},{"comment":"The text refers to 'Fig. 2B', but the figure panels are not lettered. Please add panel labels (a) and (b) for the ROCS-TC and docking rows, or reword the reference.","section":"Figure 2 and text around it"},{"comment":"The early-termination condition 'if n<N and ρ≤ρ_max' is not fully consistent with the while-loop pseudo-code that uses an undefined variable 'count'. Please clarify how ρ_max and the sampling rejection threshold interact, and remove the undefined variable.","section":"Appendix A.1, Algorithm 1"},{"comment":"The sentence '10K objective function calls were again budgeted for each of 10 rounds' is ambiguous about the total budget; state explicitly that SALSA uses 100K calls total, matching the LibINVENT allocation.","section":"Section 3, Multi-parameter objectives"},{"comment":"The docking score normalization is described as dividing by 'a factor of -24', -17, and -24; please clarify whether this means multiplying by -1/24 etc., and state the sign convention (whether lower raw docking scores are better).","section":"Appendix A.7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a workshop paper that addresses a practically relevant problem. The central idea is sound and the enumerated 1M-space benchmark is a strong piece of evidence. However, the load-bearing additivity assumption is only tested on smooth, globally additive-like objectives, and the scaling and ADMET claims are less controlled. For a journal version, I would expect either an explicit interaction-dominated objective experiment or a clear scope restriction, plus proper error reporting. No citation or novelty concerns beyond the standard expectation that code/data would be released for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: SALSA is a real algorithmic idea — factor active learning over synthon sets rather than full molecules — and it comes with the right kind of evidence: an exhaustively enumerated 1M-molecule ground truth, a strong MolPAL baseline, and a clean story. The main soft spot is exactly the one the authors admit: the additive/synthon-independence assumption is only tested on two scoring functions, both likely near-additive for the chosen core. I'd want to see an interaction-dominated objective before trusting the general claim.\n\nWhat's new: each R-group position is treated as a separate bandit, with per-vector MPNN surrogates trained on synthon-score pairs, and top-scoring synthons are combined to make molecules. This changes inference cost from O(product) to O(sum), enabling screening of ~2T molecules. The 1M-space benchmark is honest: full enumeration and scoring, so recall of top-1K is a real number. SALSA achieves 96.5% and 94.5% recall versus MolPAL's 98.5% and 95.4% — a small, acceptable degradation. The tabular Thompson sampling comparison shows the value of generalizing across synthons. The MPO experiments against LibINVENT use equal budgets (100K calls), with SALSA showing more scaffold diversity.\n\nSoft spots, in proportion:\n\n- The independence assumption is load-bearing. The surrogate assigns each molecule's full score to both constituent synthons, so each synthon sees scores that vary with its partner. Appendix A.4 says the aleatoric variance from the unobserved complementary synthon dominates. That means SALSA's success depends on how well the objective factorizes. ROCS-TC is known to be roughly additive; docking is less so but may still work for this core. There's no test with a deliberately interaction-heavy objective, e.g. a pharmacophore requiring two specific R-groups or a pairwise penalty. The MPO functions are just linear combinations of the same two scores plus QED, so they add no evidence about robustness. The authors acknowledge this in the conclusion, which is good, but the gap remains.\n\n- Fig. 2 shows curves smoothed over 5 trials without error bars. Minor for a workshop paper.\n\n- No code, data, or SMIRKS definitions released. Exact reproduction is impossible. This is common at workshops, but worth noting.\n\n- The ADMET comparison uses predicted values from internal models, not measurements. The \"comparable to known bioactives\" claim should be read with that in mind.\n\n- Scaling experiments lack ground truth; the 2T result shows SALSA finds better scores as space grows, but there's no strong baseline beyond its own trajectory.\n\nOverall, the algorithmic contribution is sound and the 1M benchmark is about as honest as it gets. The independence assumption is a real limitation, but it's clearly stated, and the method works on the tested objectives. I'd send this to peer review — it deserves referee time — with a request for an interaction-dominated experiment and code release.\n\nThis is for anyone working on ultra-large library screening, active learning for molecular design, or reaction-based combinatorial spaces. I'd cite it in my own virtual-screening work.","headline":"Factored synthon-level active learning is a genuine extension of pool-based AL to non-enumerable spaces, with an honest 1M ground-truth benchmark; the main gap is the narrowly tested independence assumption.","tokens_in":12590,"tokens_out":4084,"would_cite":true,"duration_ms":38973,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By scoring fragments instead of whole molecules, active learning screens trillion-compound spaces and finds 95% of top hits early.","keywords":["active learning","synthon","multi-vector expansion","virtual screening","Thompson sampling","ultra-large chemical spaces","molecular design","multi-parameter optimization"],"falsifier":"Take a small enumerated space and define a score that is the sum of per-synthon terms plus a large cooperation bonus when two particular synthons co-occur; run SALSA on this objective and check whether recall of the true top-1,000 combinations falls far below the 95% observed on additive objectives.","tokens_in":11563,"feed_emoji":"🧪","tokens_out":6261,"duration_ms":60723,"temperature":0.7,"pith_summary":"The paper tries to extend pool-based active learning to chemical libraries too large to enumerate. Its proposal is to model and acquire one fragment at each reaction site, treating each site as an independent bandit, so the cost of ranking a space grows with the number of fragments rather than the number of molecules. On a benchmark of one million molecules, this factored approach recovers about 95% of the best thousand compounds after scoring a small fraction of the space, and it scales to roughly two trillion compounds. The authors also show that molecules selected this way match or exceed a generative baseline on multi-objective design tasks for three protein targets while producing more diverse scaffolds.","feed_headline":"Fragment-level active learning finds 95% of top hits fast","feed_subtitle":"SALSA screens up to ~2 trillion compounds and beats generative design on multi-objective tasks.","key_machinery":"The load-bearing object is the factored synthon-level surrogate model with independent acquisition. A synthon is a fragment or R-group attached at one vector position; the paper treats each position as a separate multi-armed bandit whose arms are synthons. A message-passing neural network predicts a mean and variance for each synthon's contribution to the molecular score, acquisition samples from those Gaussian predictions via Thompson sampling, and the top synthons across vectors are combined into candidate molecules. This factorization, reducing a product of pool sizes to a sum, is what makes non-enumerable spaces tractable; the roughly additive structure of shape-based scores is what makes the factorization accurate enough to rank well.","core_discovery":"The central claim is that multi-vector molecular expansion can be screened without enumerating the product space. The algorithm attributes each scored molecule's objective value to its constituent synthons, trains a surrogate per vector, and uses Thompson sampling to pick synthons whose predicted score distributions look promising; assembling those synthons yields the next batch of molecules. This changes inference-time complexity from the product of synthon-pool sizes to their sum, so spaces with trillions of combinations become practical. In an enumerated one-million-molecule space the method identifies 96.5% of the top-1,000 compounds for a shape-based objective and 94.5% for docking, close to full-molecular active learning, and its top-scoring molecules improve as the space is enlarged to roughly 910K by 2.4M synthons, about two trillion molecules. For multi-parameter objectives combining docking or shape similarity with drug-likeness, the method produces molecules with drug-like property distributions and greater scaffold diversity than a generative baseline.","pith_inferences":["If objective scores contain strong interactions between specific R-groups, such as cooperative binding or steric clashes between two substituents, the independent-synthon surrogate can misrank combinations; testing SALSA on a deliberately coupled objective would quantify this failure mode.","Because acquisition is linear in the number of synthons, the same factoring could extend to more than two vectors or to scaffold hopping by treating the core as an additional synthon vector.","The per-synthon score distributions learned by SALSA could be reused to estimate full-library score distributions, guiding which sub-libraries to enumerate rather than only which molecules to sample.","The method's success on shape-based scores suggests fragment-additive objectives are the natural first targets; structure-based scores that depend on global geometry may need joint models to reach the same efficiency."],"forward_implications":["Ultra-large make-on-demand libraries, with billions to trillions of compounds, become accessible to active learning with expensive scoring functions such as docking.","Every candidate molecule comes with an explicit synthesis route, because each synthon is tied to a reaction; this eases the design-to-make transition.","Practitioners can steer the search by filtering synthon sets for physicochemical properties or practical procurement concerns before screening begins.","Factored acquisition degrades minimally relative to full-molecular active learning on objectives that are roughly additive over fragments, and can learn faster in early rounds.","On multi-parameter objectives, SALSA matches or improves over a generative approach while yielding substantially more unique high-scoring scaffolds."],"supporting_citations":[{"why":"Supplies the pool-based active learning framework and Thompson sampling strategy that SALSA adapts to multi-vector spaces, and the MolPAL baseline it must match.","marker":"Graff et al. (2021)"},{"why":"Provides the tabular Thompson sampling baseline demonstrating the need for generalization across synthons.","marker":"Klarich et al. (2024)"},{"why":"Supplies the message-passing neural network implementation used as the synthon surrogate model.","marker":"Heid et al. (2024)"},{"why":"The reaction-based generative method used as the multi-vector design baseline in the multi-parameter objective experiments.","marker":"Fialková et al. (2021)"},{"why":"Provides the implementation framework and configuration for the generative baseline.","marker":"Loeffler et al. (2024)"},{"why":"Defines the synthon construction trick of replacing displaced reactive groups with a generic linker atom, used to build the search space.","marker":"Liphardt & Sander (2023)"},{"why":"Demonstrates the scale of synthon-based virtual libraries that SALSA targets.","marker":"Sadybekov et al. (2022)"}],"fun_headline_variants":["SALSA learns synthon scores to screen trillions of compounds","Fragment-based active learning finds 95% of top hits","Synthon-level active learning scales to trillion-compound spaces","SALSA: active learning on synthons screens trillions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes a molecule's objective score is close to the sum of independent contributions from its R-groups, so that a model trained on individual synthons can rank full molecules without ever seeing them together.","fun_headline_variants_meta":{"raw":{"variants":["SALSA learns synthon scores to screen trillions of compounds","Fragment-based active learning finds 95% of top hits","Synthon-level active learning scales to trillion-compound spaces","SALSA: active learning on synthons screens trillions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3061,"prompt_tokens":886,"completion_tokens":2175,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":2103}},"tokens_in":502,"tokens_out":2175,"duration_ms":17493,"temperature":1.0,"reasoning_tokens":2103,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:23:40.236699+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small enumerated space and define a score that is the sum of per-synthon terms plus a large cooperation bonus when two particular synthons co-occur; run SALSA on this objective and check whether recall of the true top-1,000 combinations falls far below the 95% observed on additive objectives.","supporting_citations":[{"cited_title":"Fast substructure search in combinatorial library spaces","cited_arxiv_id":null,"evidence_quote":"Defines the synthon construction trick of replacing displaced reactive groups with a generic linker atom, used to build the search space."}],"review_version":1}